From: Konstantinos Margaritis <konstantinos@vectorcamp.gr>
Date: Tue, 20 Jul 2021 11:38:19 +0000 (+0000)
Subject: fix arm loadu_maskz()
X-Git-Tag: v5.4.3+vectorscan~69
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=825460856f750d067063e4fb54f5949f2b53be84;p=thirdparty%2Fvectorscan.git

fix arm loadu_maskz()
---

diff --git a/src/util/supervector/arch/arm/impl.cpp b/src/util/supervector/arch/arm/impl.cpp
index 678ebdfb..2c4cf3d6 100644
--- a/src/util/supervector/arch/arm/impl.cpp
+++ b/src/util/supervector/arch/arm/impl.cpp
@@ -289,11 +289,11 @@ really_inline SuperVector<16> SuperVector<16>::load(void const *ptr)
 template <>
 really_inline SuperVector<16> SuperVector<16>::loadu_maskz(void const *ptr, uint8_t const len)
 {
-    uint8_t alignment = (uintptr_t)(ptr) & 15;
-    SuperVector<16> maskb = Ones() << alignment;
-    SuperVector<16> maske = Ones() >> (16 -len - alignment);
-    SuperVector<16> v = SuperVector<16>::loadu((const m128 *)ptr);
-    return {maskb.u.v128[0] & maske.u.v128[0] & v.u.v128[0]};
+    SuperVector<16> mask = Ones() >> (16 -len);
+    mask.print8("mask");
+    SuperVector<16> v = loadu(ptr);
+    v.print8("v");
+    return mask & v;
 }
 
 #ifdef HS_OPTIMIZE